[vLLM IR] Add IR op testing and benchmarking infrastructure (#40167)

Signed-off-by: Yanan Cao <gmagogsfm@gmail.com>
Co-authored-by: Theresa Shan <Theresa.Shan@amd.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Yanan Cao
2026-04-21 00:23:03 +00:00
committed by GitHub
co-authored by Theresa Shan Claude Opus 4.6
parent 6867bcd076
commit fe5c115ee4
12 changed files with 674 additions and 46 deletions
+46
View File
@@ -0,0 +1,46 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""
Shared test utilities for vLLM IR op correctness tests.
"""
import torch
from vllm.ir.op import IrOp
NUM_TOKENS = [1, 8, 17, 32, 512, 2048]
COMMON_HIDDEN_SIZES = [
2048, # Llama 3.2 1B, Qwen 3 MoE 30B-A3B, Gemma 3n
4096, # Llama 3 8B, Qwen 3 8B
5120, # Llama 4 Scout 17B-16E
7168, # DeepSeek V3
8192, # Llama 3 70B
]
def clone_args(args: tuple) -> tuple:
return tuple(a.clone() if isinstance(a, torch.Tensor) else a for a in args)
def supported_providers(op: IrOp) -> list[str]:
return [
name for name, impl in op.impls.items() if name != "native" and impl.supported
]
def assert_close(op: IrOp, actual, expected):
if isinstance(actual, torch.Tensor):
tol = op.get_tolerance(actual.dtype)
try:
torch.testing.assert_close(actual, expected, **tol)
except AssertionError as e:
raise AssertionError(
f"{e}\n\nTo adjust tolerance, use:\n"
f" ir.ops.{op.name}.override_tolerance("
f"{actual.dtype}, atol=..., rtol=...)"
) from None
elif isinstance(actual, (tuple, list)):
for a, ex in zip(actual, expected):
assert_close(op, a, ex)
else:
assert actual == expected
+65
View File
@@ -495,3 +495,68 @@ def test_uuid_and_oot(tmp_path: Path):
assert uuid2 == uuid
assert uuid2 != uuid1
del _custom_mm.impls["impl_mm_oot"]
def _test_native(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
return x + y
def _make_op_with_generator(name: str = "_ig_test"):
op = IrOp(name, _test_native)
@op.register_input_generator
def _gen(n: int = 4):
x = torch.randn(n, 3)
y = torch.randn(n, 3)
return x, y
return op
def _test_native_single(x: torch.Tensor) -> torch.Tensor:
return x
class TestInputGenerator:
def test_no_input_generator_by_default(self):
op = IrOp("_ig_test_no_gen", _test_native_single)
assert not op.has_input_generator
def test_register_input_generator(self):
op = _make_op_with_generator("_ig_test_reg")
assert op.has_input_generator
def test_generate_inputs_returns_tuple(self):
op = _make_op_with_generator("_ig_test_tuple")
result = op.generate_inputs(n=2)
assert isinstance(result, tuple)
assert len(result) == 2
assert result[0].shape == (2, 3)
assert result[1].shape == (2, 3)
def test_generate_inputs_default_kwargs(self):
op = _make_op_with_generator("_ig_test_default")
result = op.generate_inputs()
assert result[0].shape == (4, 3)
def test_generate_inputs_without_registration_raises(self):
op = IrOp("_ig_test_no_gen_raises", _test_native_single)
with pytest.raises(RuntimeError, match="No input generator"):
op.generate_inputs()
class TestTolerance:
def test_override_and_get_tolerance(self):
op = IrOp("_tol_test", _test_native)
tol = op.get_tolerance(torch.float32)
assert tol == {"atol": 1e-5, "rtol": 1.3e-6}
op.override_tolerance(torch.float32, atol=0.1, rtol=0.2)
assert op.get_tolerance(torch.float32) == {"atol": 0.1, "rtol": 0.2}
assert op.get_tolerance(torch.float16) == {"atol": 1e-3, "rtol": 1e-3}
def test_get_tolerance_raises_for_unknown_dtype(self):
op = IrOp("_tol_test_unknown", _test_native)
with pytest.raises(ValueError, match="No tolerance defined"):
op.get_tolerance(torch.complex64)